The Loser's Game: Why Software Survives AI by Not Losing

Larix Research · Framework · The businesses named below are large and well-covered; they are illustrations, not coverage. No rating, no target. Disclosures at the end.

A game amateurs lose and professionals win

In the 1970s the scientist Simon Ramo noticed something about tennis that turned out to be about far more than tennis. Professional tennis is a winner's game: the outcome is decided by the winner's decisive, intentional shots. Amateur tennis is a loser's game: the outcome is decided by what the loser does. Points are not won with brilliance; they are given away with unforced errors — balls into the net, double faults, shots sent long. Ramo estimated that amateurs surrender roughly 80% of points rather than earn them.

In 1975 Charles Ellis pointed that observation at institutional investing and concluded that money management had become a loser's game, in which the disciplined path to winning was to stop making errors rather than to chase brilliance. The idea became the intellectual foundation of indexing, and it rests on a single question: is this a game you win by hitting winners, or by not missing? Get the classification wrong and every subsequent decision is wrong.

We think the software industry is being misclassified in exactly this way right now, and that the misclassification is the most exploitable analytical error in the market.

What AI actually deflates

The 2025–26 repricing of software has a coherent thesis underneath it. Application software is valued on the durability of its revenue, and generative AI attacks durability from three directions: it collapses the cost of writing replacement software, it lets customers build internally what they once licensed, and agentic systems threaten to execute workflows directly, demoting the seat-based tool that used to rent you that workflow. The market looked at this and repriced the entire category, halving multiples across names whose revenue had not fallen at all.

The thesis is right that AI deflates capability — the ability to perform a task in software. When the marginal cost of producing working software approaches zero, the scarcity value of any given tool approaches zero with it. Businesses whose entire product is capability are correctly frightened. The carelessness is in assuming every business reported in the software column sells capability. Many of the most valuable ones do not. They sell something AI cannot manufacture — and in several cases, something AI makes structurally more valuable. In Ramo's terms: the market assumed everyone was playing a winner's game, when the most defensible players were quietly winning a loser's game.

The interface is the winner's shot; the data is the ball in play

In an AI-disrupted stack, the interface is a winner's shot and the data underneath is the ball in play. Interfaces — the search box, the dashboard, the app — are exactly what agentic AI commoditises. If an assistant can query a system and present the answer inside a chat window, the elaborately designed front end becomes a layer the agent routes around. Spending to build a more brilliant interface, in that world, is trying to hit winners in a game that no longer rewards them.

But the agent has to query something. If that something is proprietary — a licensed real-time data feed, decades of accumulated transaction history, a settlement rail, a contractual supply relationship — then the interface commoditising above it does not weaken the asset. It concentrates value onto it. The agent that replaces your dashboard still has to source the data your dashboard displayed, and if you own that data under commercial terms, the agent is not your disintermediation. It is your next customer.

This reframes what defensibility means in the AI era. It is not the sophistication of your software — that is the depreciating asset. It is whether, when every interface in your category becomes free, there is a proprietary substrate the free interfaces are forced to pay to reach.

The exchange test

The cleanest illustration is the business that most resembles a financial exchange, because exchanges have already lived through this and demonstrate the equilibrium.

Consider a sports-data company — Sportradar is the scaled example — often described as analytics and betting software. Described that way, it looks acutely AI-exposed: surely an assistant will simply answer "what's the live win probability" directly. But that framing confuses the interface with the asset. The asset is the licensed, rights-locked, real-time feed — the officially sanctioned record of what is happening on the field. If consumption shifts entirely to "ask an AI," the AI still has to source the live data from the entity that holds the rights to it. The interface changed; the toll did not move. Deutsche Börse occupies the same position with respect to the DAX: whoever builds the app, the chart, or the chatbot that displays the index, the number itself is licensed from one place. An exchange does not fear a new order-entry client. It meters it.

A classifieds marketplace passes the same test with a different substrate — liquidity and supplier contracts rather than a data licence. The agent that wants the inventory must query the party that owns it, on that party's terms. The AI-era question is never "can an agent replace the interface" — it can — but "when it does, does it commoditise you, or does it become one more metered client of an asset it cannot reproduce."

Scout24 is the named case in our own coverage region, and its numbers show the loser's-game playbook run deliberately. ImmoScout24's asset is the deepest pool of property inventory in Germany and the agent relationships behind it; no assistant layer can conjure that liquidity, only query it. Group EBITDA margin expanded to 60.1% in early 2026, with analysts underwriting a path toward 61% — leverage that management attributes to a product-led strategy that integrates AI into the core platform rather than chasing a standalone AI product, while higher-value private subscription tiers monetise the liquidity. Capability got cheaper; the ball in play got more profitable. That is the trade the market's blanket discount has backwards.

The test generalises beyond data licences — and beyond finance. Wolters Kluwer's clinical decision product UpToDate was marked down on the fear that a physician will simply ask a general assistant. But what a hospital buys is not an interface to medical text; it is an authoritative, liability-bearing corpus — evidence graded and maintained by thousands of named clinicians, citable in a way a chatbot's synthesis is not. An assistant that wants to practise medicine does not compete with that corpus; it licenses it. The product was never information retrieval — it was accountability. And the equilibrium is no longer hypothetical: Wolters Kluwer has expanded an enterprise AI collaboration with OpenAI — the assistant layer contracting with the corpus rather than around it — while shipping expert AI inside its own regulated platforms and cloud products such as its Genya suite for Italian accountants. Recent analyst upgrades cite exactly this strength; that is the framework's claim restated in sell-side language.

The pattern is the same everywhere: somewhere in the stack sits a regulated, contractual, or accumulated asset that the free interfaces above are forced to transact with, on terms the asset owner writes.

The unforced error is inaction

Here the framework turns from valuation to management, and it resolves a tension: the loser's-game frame counsels against swinging for winners, yet we are about to argue these companies should adopt AI aggressively. The resolution is the point. In this era, the unforced error is inaction. Adopting AI inside the business — above all in how software itself is built — is not going for a winner. It is the new definition of keeping the ball in play. The double fault of the current moment is watching a competitor collapse its development cost and cycle time with agentic engineering while you leave adoption to the enthusiasm of individual teams.

What would be swinging for a winner — the genuinely dangerous, low-percentage shot — is betting the company on a proprietary frontier AI product, a house model, a moonshot feature meant to leapfrog the category. In a loser's game, most who attempt it hit the ball long. The disciplined posture separates the two cleanly: adopt AI ruthlessly in how you build and operate, because that is error-avoidance and the table stakes of staying in the game; do not bet the franchise on building the AI product that saves you.

This is why the most consequential AI decision at a defensible software business is usually not in the product at all. It is in the software factory — standardising development itself on AI agents, with engineers moving to specification, review and orchestration. A company sitting on a proprietary data or liquidity moat, whose input cost is the software it builds every year, is on the winning side of the exact trade the market punished: AI deflates the price of the capability it buys, while leaving the asset it sells untouched. The error is failing to act like it.

The discipline

Stripped to its instruction, the framework is short. Ask Ellis's question of each name: is this a winner's game or a loser's game — and is the ball it keeps in play proprietary or ambient? The businesses punished as capability sellers that are in fact liquidity or data franchises are misclassified, and misclassification is where mispricing lives. Their task is not to dazzle. It is to protect the proprietary substrate the interface layer will be forced to pay to reach, and to commit the one act of aggression the loser's game actually rewards — pointing AI at their own cost of building software before a competitor does. In 2026, the net you must clear is your own rate of adoption.

The companies that internalise this keep their franchises and compound. The ones that mistake it for a winner's game — and spend the next three years chasing brilliant AI products instead of defending their data and rebuilding their software factory — will lose the point the way amateurs always have. Not because the other side hit a winner. Because they hit it long.

Sources and method

This is a framework note, not an initiation of coverage, and contains no recommendation, rating, or price target. The winner's/loser's-game distinction originates with Simon Ramo's observations on tennis and was applied to investment management by Charles D. Ellis in Winning the Loser's Game (1975); the application to AI-era software strategy is our own. Companies named (Sportradar; classifieds marketplaces including SMG Swiss Marketplace Group and Scout24; Deutsche Börse; Wolters Kluwer) appear as illustrations of a structural argument and are drawn from public reporting; several are covered by sell-side research and sit outside the undercovered universe this firm exists to examine. We hold no position in the securities mentioned and received no compensation from any party in connection with this note.